To Index or Not to Index: Optimizing Maximum Inner Product Search
share
Making top-K predictions for state-of-the-art Matrix Factorization models requires solving the Maximum Inner Product Search problem. Solving MIPS can be computationally expensive, thus spurring the recent development of several indexing techniques for this task. These techniques generally exploit similarity between user or item weights in the models to accelerate predictions. In this paper, we show that the current state of the art does not always outperform brute-force matrix multiplication: these models may have significantly less similarity than can be exploited by these techniques. To address this problem, we propose RecOpt, a system that uses an efficient, sampling-based estimation technique to automatically choose between indexing or brute force. In addition, we propose a new baseline indexing scheme, RecDex, that can leverage blocked linear algebra to improve indexing-based serving performance. Together, RecOpt and RecDex outperform state-of-the-art indexes by 3.2× on average, and up to 10.9×, on widely studied models for recommendations and MIPS.
READ FULL TEXT